Deep Learning to Predict Traumatic Brain Injury Outcomes in the Low-Resource Setting

Area of Science:

• Neurology
• Artificial Intelligence
• Global Health

Background:

• Traumatic brain injury (TBI) significantly impacts low- and middle-income countries (LMICs), where accurate patient prognostication is challenging but crucial for effective care.
• Limited resources in LMICs necessitate efficient patient triage and management strategies for TBI patients.

Purpose of the Study:

• To develop and evaluate the first deep learning model for predicting TBI patient outcomes.
• To compare the performance of deep learning models against less complex algorithms for TBI prognostication.
• To enhance TBI triage and decision support in resource-limited settings.

Main Methods:

• Prospective collection of TBI patient data in Kampala, Uganda (2016-2020).
• Development of deep neural network, shallow neural network, and elastic-net regularized logistic regression models.
• Utilized 13 easily acquirable clinical variables and 5-fold cross-validation for performance assessment, including AUC and AUPRC.

Main Results:

• A deep neural network model achieved the highest Area Under the Receiver Operating Characteristic Curve (0.941).
• The shallow neural network model yielded the best Area Under the Precision-Recall Curve (0.770).
• Elastic-net regularized logistic regression demonstrated non-inferior performance in several comparisons.

Conclusions:

• This study introduces the first application of deep learning for TBI prognostication, specifically targeting LMICs.
• The choice of algorithm for TBI outcome prediction should be tailored to the specific clinical context.
• Deep learning offers potential as a valuable tool for TBI decision support in resource-constrained environments.